LIMITS AND APPROXIMATIONS FOR THE M/G/1 LIFO WAITING-TIME DISTRIBUTION

Transcription

1 LIMITS AND APPROXIMATIONS FOR THE M/G/1 LIFO WAITING-TIME DISTRIBUTION by Joseph Abate 1 and Ward Whitt 2 April 15, 1996 Revision: January 2, 1997 Operations Research Letters 20 (1997) Hammond Road, Ridgewood, NJ AT&T Research, Room 2C-178, Murray Hill, NJ (wow@research.att.com)

2 Abstract We provide additional descriptions of the steady-state waiting-time distribution in the M/G/1 queue with the last-in first-out (LIFO) service discipline. We establish heavy-traffic limits for both the cumulative distribution function (cdf) and the moments. We develop an approximation for the cdf that is asymptotically correct both as the traffic intensity ρ 1 for each time t and as t for each ρ. We show that in heavy traffic the LIFO moments are related to the FIFO moments by the Catalan numbers. We also develop a new recursive algorithm for computing the moments. Key Words: M/G/1 queue; Last-in first-out service discipline; Waiting time; Heavy Traffic; Asymptotics; Asymptotic normal approximation; Moments; Catalan numbers

3 1. Introduction The purpose of this paper is to deduce additional properties of the steady-state waiting-time distribution (until beginning service) in the M/G/1 queue with the last-in first-out (LIFO) service discipline. We obtain our results here by applying results in our previous papers [1] [6]. Our results here complement previous M/G/1 LIFO results by Vaulot [17], Wishart [19], Riordan [13], [14], Takács [16] and Iliadis and Fuhrmann [10]. We start in Section 2 by establishing a heavy-traffic limit theorem for the M/G/1 LIFO steadystate waiting-time distribution. In Section 3 we give a new derivation for the large-time asymptotics of the LIFO steady-state waiting-time distribution, complementing our recent result in [1]. In Section 4 we then develop an approximation for the LIFO waiting-time distribution, called the asymptotic normal approximation, and show that it is asymptotically correct both as time t for each value of the traffic intensity ρ and as ρ 1 for each t. Our approximation generalizes an approximation suggested for the M/M/1 model by Riordan [14], p Sections 2 4 parallel and draw on our previous limiting results for the M/G/1 busy-period distribution in [5]. In Section 5 we give a numerical example showing that the asymptotic normal approximation for the LIFO steadystate waiting-time cdf performs very well. We show that it is much better than the large-time limit (the limit as t for each fixed ρ). In Section 6 we develop a new recursive algorithm for computing the M/G/1 LIFO waitingtime moments. We believe that this algorithm is an attractive alternative to previous algorithms developed by Takács [16] and Iliadis and Fuhrmann [10]. The first four moments are given explicitly. In Section 7 we show that the moments have a very simple asymptotic form in heavy traffic. 2. Heavy-Traffic Limit for the LIFO CDF In this section we establish a heavy-traffic limit for the LIFO steady-state waiting-time cdf, denoted by W L. For this purpose, we establish heavy-traffic limits for the M/G/1 workload process moment cdf s. These workload results extend previous results for the M/M/1 special case in Corollary 5.22 of [3]. The heavy-traffic limits for the workload process are local-limit refinements (versions for densities) to theorems that can be deduced directly from older heavy-traffic limit theorems, e.g., in Whitt [18]. We use the following notation. For any cumulative distribution function (cdf) F, let f be the associated probability density function (pdf) and let m k (F ) be the k th moment. Let F c (t) 1 F (t) 1

4 be the associated complementary cdf (ccdf) and let the associated equilibrium excess ccdf be Fe c (t) = m 1(F ) 1 F c (u). (2.1) The pdf of F e is f e and its k th moment is m k (F e ). For k 1, let H k be the k th moment cdf of canonical (drift 1, diffusion coefficient 1) reflected Brownian motion (RBM) {R(t) : t 0}, i.e., t H k (t) = E[R(t)k R(0) = 0] E[R( ) k ], t 0. (2.2) By definition, H k (t) as a function of t is the expectation of a stochastic process, but it also has the structure of a cdf. Throughout this paper we relate the tail behavior of cdf s to large-time asymptotics of stochastic processes. Properties of the RBM moment cdf s are established in [2]. Let h k be the density of H k. Let Φ be the standard (mean 0, variance 1) normal cdf and let φ be its density. We shall primarily focus on the pdf h 1, which has the formula h 1 (t) = 2t 1/2 φ(t 1/2 ) 2[1 Φ(t 1/2 )] = 2γ(t) γ e (t), (2.3) where γ is the gamma pdf with mean 1 and shape parameter 1/2, i.e., γ(t) = (2πt) 1/2 e t/2, t 0, (2.4) and γ e is the associated equilibrium-excess pdf. We consider a family of M/G/1 queueing systems indexed by the arrival rate ρ. We assume that the service-time distribution is fixed, having cdf G with pdf g, mean 1 and second moment m 2 (G). Let {V ρ (t) : t 0} be the M/G/1 workload (unfinished work or virtual waiting time) process and, for k 1, let H ρk be its associated k th moment cdf, expressed as a function of the arrival rate ρ, i.e., H ρk (t) = E[V ρ(t) k V ρ (0) = 0] E[V ρ ( ) k ], t 0, (2.5) see [4]. Just as with H k (t) in (2.2), H ρk (t) as a function of t is simultaneously the expectation of a stochastic process and a cdf for each ρ. Also let H ρ0 be the 0 th -moment cdf or server-occupancy cdf, defined by H ρ0 (t) = (1 P 00 (t))/ρ, t 0, (2.6) where P 00 (t) is the probability of emptiness at time t starting empty at time 0 (with arrival rate ρ), as in (23) of [4]. 2

5 The key formula connecting these expressions to the LIFO cdf W L is WL c (t) = P 00(t) (1 ρ) = ρhρ0 c (t), (2.7) due to Takács [16]; see (36) of [1] and (78) on p. 500 of [16]. The next theorem describes the LIFO steady-state waiting-time cdf W L in heavy traffic. Theorem 2.1 For each t > 0, for h 1 in (2.3). lim ρ 1 2(1 ρ) 1 W c L (tm 2(G)(1 ρ) 2 ) = lim ρ 1 2(1 ρ) 1 H c ρ0 (tm 2(G)(1 ρ) 2 ) = lim ρ 1 m 2 (G)(1 ρ) 2 h ρ1 (tm 2 (G)(1 ρ) 2 ) = h 1 (t) Proof. By (2.7), the first limit is equivalent to the second. By Theorem 2(b) of [4], h ρ0e (t) = H ρ1 (t), so that h ρ1 (t) = H ρ0 (t)/h ρ01. By Theorem 6(b) of [4], h ρ01 = ρ 1 v ρ1 = m 2 (G)/2ρ(1 ρ). Hence the third limit is equivalent to the second. By Theorems 3(a) and 4(b) of [4], ĥ ρ1 (s) = 1 ˆb ρe (s) sb ρe1 (1 ρ + ρˆb ρe (s)), (2.8) where ˆb ρɛ (s) is the transform of the equilibrium excess distribution associated with the busy period and b ρe1 is its mean. For brevity, let c = (1 ρ) 2 /m 2 (G) and note that b ρe1 = (2c) 1 by Theorem 5(b) of [4]. Then, by (2.8), ĥ ρ1 (cs) = 2[1 ˆb ρe (cs)] s(1 ρ + ρˆb ρe (cs)). Since we have expressions for the transforms, we establish the convergence using them. By Theorem 1 of [5], ˆb ρe (cs) ĥ1(s) as ρ 1. (Equations (2.9) of [5] should read h ρ (t) m 2 (1 ρ) 2 b ρe (tm 2 (1 ρ) 2 ) = m 2 (1 ρ) 1 B c ρ (tm 2(1 e) 2 ), with m 2 there being m 2 (G) here. Incidentally, a space normalization typically c(1 ρ) for a constant c is missing before W ρ (dt(1 ρ) 2 ) in Condition C2 on p. 589 in [5] as well.) Hence, 2[1 ĥ1(s)] ĥ ρ1 (cs) sĥ1(s) with the last step following from (2.6) of [5] or p. 568 of [2]. The LIFO approximation obtained from Theorem 2.1 is W c L = ĥ1e(s) ĥ 1 (s) = ĥ1(s), (1 ρ) (t) h 1 (t(1 ρ) 2 /m 2 (G)). (2.9) 2 3

6 The M/M/1 special case of Theorem 2.1 for h ρ1 (t) is given in Corollary of [3]. In [3] the time scaling is done at the outset; see (2.2) there. The limit for the density h ρ1 implies a corresponding limit for the associated cdf H ρ1, by Scheffé s theorem, p. 224 of Billingsley [7]. The following result could also be deduced from standard heavytraffic limits for the workload process [18]. Corollary. For each t > 0, lim H ρ1(tm 2 (G)(1 ρ) 2 ) = H 1 (t) = 1 2(1 + t)[1 Φ(t 1/2 )] + 2t 1/2 φ(t 1/2 ). ρ 1 3. Large-Time Asymptotics In [1] we used (2.7) and an integral representation for P 00 (t) to establish the asymptotic behavior of WL c (t) as t. Here we give an alternative derivation based on asymptotics for the density w L (t). Let f(t) g(t) as t mean that f(t)/g(t) 1 as t. We start with the elementary observation that the conditional LIFO steady-state waiting time given that it is positive coincides with the busy period generated by the equilibrium excess of a service time. Let b(t, θ) be the density of a busy period starting with a service time of length θ. Then the LIFO waiting-time distribution has an atom of size 1 ρ at the origin and a density w L (t) = ρ b(t, θ)g e (θ)dθ, t > 0. (3.1) 0 From (3.1), we reason as in Cox and Smith [8] and [1] to obtain the following asymptotic result. Theorem 3.1 As t, and w L (t) (ω/τ)(πt 3 ) 1/2 e t/τ (3.2) W c L (t) ω(πt3 ) 1/2 e t/τ, (3.3) where τ 1 is the asymptotic decay rate (reciprocal of the relaxation time), given by τ 1 = ρ + ζ ρĝ( ζ), (3.4) ζ is the unique real number u to the left of all singularities of the service-time moment generating function ĝ( s) (assumed to exist) such that ĝ ( u) = ρ 1, (3.5) 4

7 ω = ρα/ζ 2, (3.6) α = [2ρ 3 ĝ ( ζ)] 1/2. (3.7) Proof. We apply (19) of [1] with (3.1) to obtain the integral representation w L (t) = (ρ/t)l 1 ( ĝ e(s) exp( ρt(1 ĝ(s))), (3.8) where L 1 is the Laplace transform inversion operator (Bromwich contour integral) L 1 (ĝ(s))(t) = 1 2πi a+i a i e st ˆf(s)ds, (3.9) where the contour Re(s) = a is to the right of all singularities of ˆf(s). We then apply Laplace s method as on p. 156 of Cox and Smith [8] and pp. 80, 121 and 127 of Olver [12] to obtain (3.2). We can also obtain the asymptotic result for w L (t) by relating (3.8) to the integral representation for the busy-period density b(t) in (3) of [1]. From p. 127 of Olver [12], we see that where so that w L (t) ρ 2 ( ĝ e( ζ)b(t) as t (3.10) w L (t) ĝ e( ζ) = 1/ζ 2 ρτ, (3.11) ρ ζ 2 b(t) as t. (3.12) τ Hence, we can invoke the known asymptotics for b(t) in (5) of [1]. Finally, we integrate (3.2) to obtain (3.3); see p. 17 of Erdélyi [9]. 4. The Asymptotic Normal Approximation In [5] we found that the tail of the M/G/1 busy-period ccdf is well approximated by an asymptotic normal approximation, which is asymptotically correct as ρ 1 for each t and as t for each ρ. We now apply essentially the same reasoning to develop a similar approximation for the LIFO steady-state waiting-time ccdf. Paralleling Theorem 2 of [5], we can express the asymptotic relation in (3.3) in terms of h 1 (t), because h 1 (t) 2t 1 γ(t) as t, (4.1) for γ in (2.4). The following is our asymptotic normal approximation, obtained by combining (2.4), (3.1) and (4.1). The M/M/1 special case was proposed for the M/M/1 queue by Riordan [14], p

8 Theorem 4.1 If (3.3) holds, then W c L (t) 2ωτ 3/2 h 1 (2t/τ) as t. for h 1 (t) in (2.3), τ in (3.4) and ω in (3.6). Theorem 4.1 directly shows that the asymptotic normal approximation is asymptotically correct as t for each ρ. Previous asymptotic relations for τ and ω show that it is also asymptotically correct as ρ 1 for each t. In particular, by (14) and (40) of [1], and τ 1 = ω = τ 2 which is consistent with (2.9). For the constant, note that 2ωτ 3/2 (1 ρ)2 (1 + O(1 ρ)) as ρ 1 (4.2) 2m 2 (G) m 2 (G) (1 + O(1 ρ)) as ρ 1, (4.3) 2 m 2 (G)/2τ 1 ρ 2 as ρ 1. (4.4) 5. A Numerical Example In this section we compare the approximations for the M/G/1 LIFO steady-state waiting-time ccdf WL c (t) to exact values obtained by numerical transform inversion. We consider the gamma service-time distribution with shape parameter 1/2, denoted by Γ 1/2, which has pdf g in (2.4) and Laplace transform Its first four moments are 1, 3, 15 and 105. ĝ(s) = 1/ 1 + 2s. (5.1) For the M/G/1 model there is no explicit result for the transform ŴL(s). One way to proceed is to determine the transform values ˆP 00 (s) from the busy period via (36) of [4] or directly via the functional equation ˆP 00 (s) = 1 s + ρĝ(1/ˆρ 00 (s)), (5.2) see (37) of [4], and then use (2.7). However, there is a more direct approach using the contour integral representation W c L(t) = t 1 L 1 (s 2 exp( ρt(1 ĝ(s)))) (1 ρ), (5.3) 6

9 where L 1 is the inverse Laplace transform operator; see (34) of [1]. Equation (3.8) above is a similar representation for the density. Table 1 gives the results for WL c (t) for the case ρ = In Table 1 we also display results for three approximations: (i) the standard asymptotic approximation in (3.1), (ii) the heavy-traffic approximation in (2.9) and (iii) the asymptotic normal approximation in Theorem 3.1. For the standard asymptotic approximation, we need to derive the asymptotic parameters ω and τ 1. For this example, we find from the root equation (3.5) that From (3.4), (3.7) and (3.6), ζ = 1 2 (1 ρ2/3 ). (5.4) τ 1 = ρ ρ2/3, (5.5) α = ρ 2/3 6 (5.6) and ω = 4ρ 1/3 6(1 ρ 2/3 ) 2. (5.7) asymptotic heavy- standard time exact by normal traffic asymptotic t numerical approximation, approximation, approximation, inversion Theorem 4.1 (2.8) (3.3) Table 1. A comparison of approximations with exact values of the LIFO steady-state waiting-time ccdf W c L (t) in the M/Γ 1/2/1 model with ρ =

10 Table 1 shows that the asymptotic normal approximation is remarkably accurate for t not too small, e.g., for t 5. Moreover, the asymptotic normal approximation is much better than the standard asymptotic approximation based on (3.3). The heavy-traffic approximation is quite good though for t neither too large nor too small, e.g., for 1.0 t 120. The standard asymptotic approximation is not good until t is very large. From Table 1, note that the standard asymptotic approximation for the tail probability is consistently high, whereas the heavy-traffic approximation (2.8) crosses the exact curve twice. 6. A Recursive Algorithm for the LIFO Moments In this section we relate the k th moments of the steady-state FIFO and LIFO waiting-time distributions, denote by v k and w k, respectively. For previous related work, see Takács [16] and Iliadis and Fuhrmann [10]. Iliadis and Fuhrmann show that the same relationship between LIFO and FIFO holds for a large class of models with Poisson arrivals. Let h 0k be the k th moment of the server-occupancy cdf H ρ0 in (2.6). By (2.7), w k = ρh 0k for all k 1. (6.1) Theorem 6 of [4] gives an expression for the moments h 0k in terms of the moments b ek of the busy-period equilibrium excess distribution, while Theorem 5 of [4] gives the recursive formula for the busy-period equilibrium-excess distribution moments b ek in terms of the moments v k. These two results give a recursive algorithm for computing the LIFO moments w k in terms of the FIFO moments v k. First, as in (43) of [4], we can relate the FIFO waiting-time moments to the service-time moments via a recursion. Let the k th service-time moment be denoted by g k. The recursion is v k = ρ 1 ρ k j=1 ( ) k gj+1 j j + 1 v k j, (6.2) where v 0 = 1, so that the FIF0 moments v k are readily available given the service-time moments g k. By Theorem 3a of [4], the equilibrium-time to emptiness transform ˆf ɛ0 (s) can be expressed as ˆf ɛ0 (s) = 1 ρ + ρˆb e (s), (6.3) so that their moments are related by f ɛ0k = ρb ek, k 1. (6.4) 8

11 Theorem 5 of [4] shows that these moments can be computed recursively, given the moments v k. To emphasize the recursive nature, for any Laplace transform ˆf(s) = let T n ( ˆf) be the truncated power series defined by where f k = 0 T n ( ˆf)(s) = 0 e st f(t)dt, (6.5) n k=0 f k k! ( s)k, (6.6) t k f(t)dt = ( 1)k d k ˆf(s) ds k s=0, (6.7) which is well defined assuming that the integrals in (6.7) are finite. Then Theorem 5(a) of [4] can be restated as f ɛ0k = k j=1 ( ) k vj j 1 ρ [T k 1( ˆf ɛ0 )] j k j (6.8) where ˆf j is understood to the j th moment, which is j!( 1) j times the j th coefficient of the power series. Then we can calculate w k from w k = ρb ek ρ k j=1 ( ) k b ej w k j, k 1, (6.9) j where w 0 = ρ, which is obtained by combining (6.1) here with Theorem 6(a) of [4]. In summary, if we want to calculate w k, then we first calculate the first k FIFO waiting-time moments v 1,..., v k recursively via (6.2). Then we calculate the first k equilibrium-time-to-emptiness moments f ɛ01,..., f ɛ0k recursively via (6.8). We obtain the associated busy-period equilibrium excess moments b e1,..., b ek via (6.4). Finally, we obtain the LIFO waiting-time moments w 1,..., w k recursively via (6.9). For example, the first four are: w 1 = v 1, w 2 = v 2 1 ρ w 3 = v 3 + 3v 2 v 1 (1 ρ) 2, w 4 = v 4 + 8v 3 v v 2 v v2 2 (1 ρ) 3, (6.10) which is consistent with Theorem 6 of [4]. In the case (e) of Theorem 6 in [4] there is a typographical error in h 04 ; all four terms there should be divided by ρ(1 ρ) 3. 9

12 7. Heavy-Traffic Limits for the LIFO Moments We now show that the LIFO waiting-time moments have a simple asymptotic form in heavy traffic. As in Section 6, let g k and w k be the k th moments of the service time and LIFO waiting time, respectively. Theorem 7.1 Assume that g n+2 <. Then w n+1 < and w n (2n 2)! (n 1)! w n 1 (1 ρ) n 1 as ρ 1. (7.1) To put Theorem 7.1 in perspective, it should be contrasted with the known heavy traffic results for first-in first-out (FIFO) and random order of service (ROS). For FIFO, the waiting-time distribution is asymptotically exponentially distributed as ρ 1, so that v n n!v n 1 as ρ 1. (7.2) For ROS, with W R as the cdf, Kingman [11] showed that 1 W R (t) 2 t/w R1 K 1 (2 t/w R1 ) where K 1 is the Bessel function, so that w Rn (n!) 2 w n R1 as ρ 1. (7.3) Of course, w 1 = w R1 = v 1. It is interesting that, as ρ 1, the n th moments v n, w RN and w n in (7.1) (7.3) are v n 1 times n!, (n!)2 and (2n 2)/(n 1)!(1 ρ) n 1, respectively. Hence, v n and w RN are O((1 ρ) n ) as ρ 1, while w n is O((1 ρ) (2n 1) ). i.e., We give another expression using the Catalan numbers. Let C n be the n th Catalan number, C n = 1 ( ) 2n ; (7.4) n + 1 n e.g., 1, 1, 2, 5, 14,... see Riordan [15]. The Catalan numbers have the self-convolution property n C n+1 = C i C n i, n 1 (7.5) i=0 The Catalan numbers are also associated with the distribution of the number of customers served in an M/M/1 busy period; see p. 65 of Riordan [14]. We combine Theorem 7.1 and (7.2) to obtain the following representation. 10

13 Corollary. Under the conditions of Theorem 7.1, n = 2. v n w n C n 1 (1 ρ) n 1 as ρ 1. The Corollary should provide a better approximation than Theorem 7.1; e.g., it is exact for Proof of Theorem 7.1. Theorem 2.1 can be interpreted as convergence of ccdf s, because h 1 (t) = h 0e (t) = 2H0 c (t). Hence we can obtain convergence of the associated moments under the regularity condition of uniform integrability; see p. 32 of [7]. The direct moment calculation from Theorem 2.1 is w n nh 1,n 1 2 n w n 1 /(1 ρ)n 1 as ρ 1, (7.6) where h 1n is the n th moments of h 1 (t) in (2.3). By (10.15) of [6], h 1n = Combining (7.6) and (7.7) gives the formula. (2n)! (n + 1)!2 n. (7.7) The extra moment condition that g n+2 be finite is used to establish the required uniform integrability. From (6.2), (6.4), (6.8) and (6.9) we see that v k, f ɛ0k, b ek and w k are all finite for k n when g k+1 <. To establish upper bounds implying uniform integrability, note that by (6.2), Using (7.8) and (6.8), v k g k+12 k (1 ρ) k. (7.8) f ɛ0k where K k is α constant independent of ρ. Finally, by (6.9) and (7.9), w k K k (1 ρ) 2k (7.9) M k, (7.10) (1 ρ) 2k 1 where M k is a constant independent of ρ. The bounds in (7.8) (7.10) provide the uniform integrability to go from convergence of distributions to associated convergence of moments. Alternatively, a proof can be based on the recursion (6.9) and asymptotics for the moments b en. By (7.2) and the recursion (6.8), From (6.9), (7.5) and (7.11), we get the conclusion. b en C nn!v1 n as ρ 1. (7.11) (1 ρ) n 11

14 References [1] J. Abate, G. L. Choudhury and W. Whitt, Calculating the M/G/1 busy-period density and LIFO waiting-time distribution by direct numerical transform inversion, Opns. Res. Letters 18, (1995). [2] J. Abate and W. Whitt, Transient behavior of regulated Brownian motion, I: starting at the origin, Adv. Appl. Prob. 19, (1987). [3] J. Abate and W. Whitt, Transient behavior of the M/M/1 queue via Laplace transforms, Adv. Appl. Prob. 20, (1988). [4] J. Abate and W. Whitt, Transient behavior of the M/G/1 workload process, Opns. Res. 42, (1994). [5] J. Abate and W. Whitt, Limits and approximations for the busy-period distribution in singleserver queues, Prob. Engr. Inf. Sci. 9, (1995). [6] J. Abate and W. Whitt, An operational calculus for probability distributions via Laplace transforms, Adv. Appl. Prob., 28, (1996). [7] P. Billingsley, Convergence of Probability Measures, Wiley, New York, [8] D. R. Cox and W. L. Smith, Queues, Methuen, London, [9] A. Erdélyi, Asymptotic Expansions, Dover, New York, [10] I. Iliadis and S. W. Fuhrmann, Moment relationships for queues with Poisson input, Queueing Systems 12, (1992). [11] J. F. C. Kingman, Queue disciplines in heavy traffic, Math. Opns. Res. 7, (1982). [12] F. W. J. Olver, Asymptotics and Special Functions, Academic Press, New York, [13] J. Riordan, Delays for last-come first-served service and the busy period, Bell System Tech. J. 40, (1961). [14] J. Riordan, Stochastic Service Systems, Wiley, New York, [15] J. Riordan, Combinatorial Identities, Wiley, New York,

15 [16] L. Takács, Delay distributions for one line with Poisson input, general holding times, and various orders of service, Bell System Tech. J. 42, (1973). [17] E. Vaulot, Delais d attente des appels telephoniques dans l ordre inverse de leur arrive, C. R. Acad. Sci. Paris 238, (1954). [18] W. Whitt, Weak convergence theorems for priority queues: preemptive-resume discipline, J. Appl. Prob. 8 (1971) [19] D. M. Wishart, Queueing systems in which the discipline is last-come first-served, Oper. Res. 8, (1960). 13